X PAPER / Full editionIssue 008

AI与科技 / Update

EmbeddingGemma 2 brings multimodal retrieval to local devices

Source author: Gorden Sun · @Gorden_Sun
Source posted:
Written and edited by

Google DeepMind’s model card clarifies the EmbeddingGemma 2 shared by Gorden Sun: it maps text, code, images, video and audio into a shared vector space for retrieval. It is an embedding model, rather than a chatbot.

The full model has 740 million parameters; the text encoder alone has 270 million. Its 768-dimensional output can be truncated to 128, 256 or 512 dimensions. The shared context limit is 8,192 tokens, including media.

The Apache 2.0 open-weight model targets local devices. It differs from the cloud-based Gemini Embedding 2 API; local use still requires suitable hardware and an indexing pipeline.

The card specifies bf16 or fp32, warning that fp16 can cause NaNs or degraded results. Queries and documents require different task prefixes. Truncated vectors need renormalization, and retrieval quality must be tested on the intended collection.

Sources and further reading

edited

Selected through a verified followed account: @Gorden_Sun.